First-Order Regret in Reinforcement Learning with Linear Function Approximation: A Robust Estimation Approach